Key takeaways
- Deliver one business domain end to end before adding more sources.
- Establish ownership, naming and access rules in the first month, not later.
- Skip exotic tooling, real-time streaming and broad self-service until the basics are trusted.
- Finish with something people use weekly, not just a loaded storage bucket.
You have approval for a data lake, a pile of source systems, and stakeholders who each expect something different. The most common failure is not technical. It is a project that ingests everything, publishes nothing and loses sponsor patience by month six. A tight 90-day plan avoids that by forcing a visible result early.
Days 1 to 30: decide, connect, secure
Pick one business domain with a clear owner and a real question, such as weekly sales by product or inventory versus demand. Confirm who will use the output and how they will judge it. Then set foundations that are cheap now and painful later.
- Define zones, for example raw, cleaned and curated, and what may live in each.
- Agree on naming and folder conventions and write them down on one page.
- Set up access by role, with sensitive data separated from the start.
- Turn on logging and cost tags so you can see who uses what.
- Connect the first two or three source systems and land raw data unchanged.
Days 31 to 60: model and test
Turn raw files into cleaned tables and a small curated model for your chosen domain. Agree on definitions with business owners for the five to ten metrics that matter, and record them in one place. Add basic data-quality checks: freshness, duplicates, missing keys and reconciliation against a source total. When a check fails, someone should be notified by name.
Days 61 to 90: publish and adopt
Connect a reporting tool to the curated layer and build the dashboard or extract that answers the original question. Walk users through it, collect feedback, and fix confusing labels. Document how data flows from source to report, who owns each step and what to do when a load fails. Finish with a short review: what did we learn about volumes, cost and skills gaps?
What to deliberately skip
- Real-time streaming, unless a named process truly needs it within minutes.
- Ingesting every table from every system just in case.
- Building a broad self-service catalog before core data is trusted.
- Custom frameworks when managed services already cover the need.
- Machine learning projects on data nobody has validated yet.
Signs you are on track
At day 90 you should be able to name the owner of each dataset, show a report that finance or operations has actually used, explain the monthly cost, and describe the next domain in line. If you have petabytes loaded but no trusted report, treat that as a warning, not a milestone.
Roles you need, even part-time
A business sponsor who can settle definitions, a data engineer to build pipelines, an analyst who knows the domain, and someone accountable for security and cost. In small teams one person may wear two hats, but each hat should have a name attached.
